The AirTrunk deal is one thread in that surge, and a useful one to pull, because it shows both how the financing works and where the strain is starting to show.
What the money buys
Blackstone is assembling a syndicate to underwrite an A$4.3 billion, roughly $3 billion, five-year loan for AirTrunk's SYD3 hyperscale data center in Australia, a facility of more than 400 megawatts. The lender list already spans Credit Agricole, DBS, Deutsche Bank, HSBC, ING, MUFG, Morgan Stanley, and United Overseas Bank, and is expected to grow.
Sit with the scale. Three billion dollars of debt for a single building, raised from a syndicate of global banks rather than a government, for a facility whose defining feature is that it consumes as much electricity as a small city. A single data center now commands financing comparable to a major piece of public infrastructure. That is not a metaphor about how big AI has become. It is the literal capital structure of one building.
And AirTrunk is doing this repeatedly. Since Blackstone and Canada's pension plan acquired it in 2024 at an enterprise value above A$24 billion, the company has committed to a $30 billion, 5-gigawatt buildout in India by 2030, a $3 billion Saudi partnership, and hyperscale projects in Malaysia and Japan, all inside about 18 months. SYD3 is the home-market piece of a global spending spree that runs on borrowed money.
The sentence in the report that matters most
Buried in the coverage is the detail that turns this from a growth story into a risk story. Banks, according to people familiar with the financing, are gradually approaching internal exposure limits for the data-center sector after committing billions to these projects.
That is the important line, and it deserves unpacking, because exposure limits are how concentration risk is supposed to be contained before it becomes a problem.
Every bank caps how much it will lend to any single sector, precisely so that a downturn in one industry cannot threaten the whole institution. When lenders start approaching those caps for data centers, three things follow in sequence. Deals get harder to syndicate, because the natural lenders are already full. Pricing rises, because scarcer lending capacity costs more. And the marginal borrower, the one with a weaker project or thinner backing, gets shut out first, while the strongest sponsors keep raising money.
AirTrunk closing this loan is therefore not evidence that the strain is imaginary. It is evidence of who still gets served when capacity tightens. A Blackstone-backed operator with a track record and hyperscale tenants is exactly the borrower banks fund last before they stop. The deals that quietly fail to happen, the ones that never make the news, are the tell, and by construction you do not see them.
Why the funding is diversifying, and what that signals
There is a second pattern worth noticing, because it points the same direction from a different angle. The borrowers are visibly hunting for new pockets of money.
In Asia-Pacific, data centers traditionally leaned on bank loans. Now the sector is spilling into other markets. Blackstone's QTS Realty recently raised a $3.25 billion leveraged loan and canceled a planned $1 billion bond sale, a leveraged loan being the kind of higher-cost debt issued by lower-rated borrowers. CoreWeave sold a $3.1 billion loan in May. Bitcoin miner TeraWulf is weighing its first leveraged loan. AirTrunk itself has explored its first data-center-backed bond.
When borrowers start reaching across bank loans, investment-grade bonds, leveraged loans, and asset-backed securities, the benign reading is diversification, spreading funding sources for resilience, which is what one MetLife executive suggested. The less benign reading is that a sector approaching the limits of its traditional lenders has to keep finding new ones to sustain the pace, and each new source tends to be more expensive or lower down the credit ladder than the last. Both readings are consistent with the facts. Which one is right depends on whether demand for the finished capacity holds up, and that is precisely the thing debt raised today is betting on.
The maturity mismatch nobody advertises
Here is the structural risk underneath all of it, and it is worth stating plainly because the deal terms make it concrete.
The AirTrunk loan is a five-year facility. The bet it funds is a multi-decade one: that AI compute demand justifies a 400-megawatt building for its entire economic life. Those horizons do not match. Five-year debt against a twenty-year asset means the loan comes due long before the building's thesis is proven, which means it has to be refinanced, probably more than once.
Refinancing is fine as long as capital stays available and cheap. It becomes the whole problem if it does not. If, when this loan matures, banks are over their exposure limits, AI demand has disappointed, or rates are higher, the refinancing that everyone is quietly assuming may be costlier or unavailable. That is the mechanism by which an infrastructure boom funded on shorter-term debt turns into a wave of distress: not because the buildings stop working, but because the debt against them comes due in a worse market than the one that issued it.
None of this means the AirTrunk loan is unsound. Blackstone is among the most sophisticated infrastructure investors in the world, AirTrunk's facilities largely serve hyperscale tenants on long contracts, and the demand for AI compute is real rather than imagined. This is a strong sponsor doing a defensible deal.
The concern is not any single loan. It is the aggregate: $334.5 billion in a single year, banks nearing their limits, borrowers fanning out into higher-cost debt markets, and short-dated loans stacked against long-dated bets, all resting on the assumption that AI demand grows fast enough, for long enough, to service and refinance the whole edifice. Each deal can be individually reasonable and the system can still be accumulating risk faster than anyone is pricing.
What to watch
Three signals will indicate which way this tips.
Whether syndication keeps working. If deals like AirTrunk's continue filling their lender groups easily, capacity is holding. If they start struggling to place debt, stretching timelines, or paying up sharply to clear, the exposure limits have started to bind in earnest.
The credit ratings on new issues. QTS's latest loan was rated Baa3, the lowest rung of investment grade. A drift of new data-center debt from investment grade toward junk would signal the quality of what is being financed is deteriorating as the volume climbs.
And any softening in the demand story, since the entire structure assumes AI compute consumption keeps growing into the capacity being built. For now the money is still flowing, and the strongest sponsors are still getting funded. The AI infrastructure boom has become one of the largest debt-financed buildouts in recent memory, and the interesting question is no longer whether the buildings get built. It is whether the borrowing that builds them can be repaid or refinanced on the timeline the lenders are assuming.